Apple Unveils M3, M3 Pro, and M3 Max: 3 Nanometer, Dynamic Caching GPU, More

I heard a good lay explanation on the most recent Upgrade podcast (around 30:00.) The explanation was provided by an anonymous Apple employee. Dynamic caching is about setting the allocated memory per GPU thread. Historically, threads received fixed allocations. These had tradeoffs since overallocation caused performance bottlenecks and right-sized allocation underutilized the RAM. Dynamic cacheing adjusts per-thread allocations so the memory per utilized thread adds up to the max available amount more often. Developers do not need to change any code to benefit.

If all that is true, it sounds like a good improvement. And it sounds plausible that Apple was the first since it would require the whole system to work in synch without introducing latency in rendering, a good use of an SoC. I’d still like to see benchmarks and understand potential failure/bottleneck scenarios.